code/slurm_logs/eval_two_selector.sbatch

#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:h200:1
#SBATCH --cpus-per-task=8
#SBATCH --mem=180G
#SBATCH --time=04:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/eval_two_selector_%j.out

set -u
cd /weka/s225250685/mats-tist
set -a; source /weka/s225250685/mats-tist/.env; set +a
export HF_HOME=/weka/s225250685/Huggingface HF_HUB_CACHE=/weka/s225250685/Huggingface/hub
export PYTHONNOUSERSITE=1 NO_PROXY=localhost,127.0.0.1 no_proxy=localhost,127.0.0.1
export PYTHONPATH=/weka/s225250685/mats-tist TOKENIZERS_PARALLELISM=false
export DB_EXEC_API_DISABLE=1

PY=/weka/s225250685/conda-envs/handbook/bin/python
VLLM=/weka/s225250685/conda-envs/handbook/bin/vllm
V6_CKPT=alignment-handbook/output/selector-qwen7b-v6-pointwise-rich
BASE_CKPT=alignment-handbook/output/sft-selector-3B-griffith-v5
DEV_FILE=eval_results/paper_SFT_VF_passAt8_bird_dev.jsonl

LOG=/weka/s225250685/mats-tist/slurm_logs/eval_two_selector_${SLURM_JOB_ID}.log
kill_vllm() { pkill -9 -f "vllm serve" 2>/dev/null || true; sleep 5; }
trap kill_vllm EXIT
wait_url() { for i in {1..360}; do curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0; sleep 5; done; return 1; }

echo "[$(date)] Serving v6 7B selector on port 8103..." | tee -a "$LOG"
$VLLM serve "$V6_CKPT" \
    --served-model-name selector_v6 --port 8103 \
    --dtype bfloat16 --gpu-memory-utilization 0.45 \
    --max-model-len 4096 --enforce-eager --disable-log-requests \
    > "$LOG.serve_v6" 2>&1 &

echo "[$(date)] Serving baseline 3B selector on port 8104..." | tee -a "$LOG"
$VLLM serve "$BASE_CKPT" \
    --served-model-name selector_base --port 8104 \
    --dtype bfloat16 --gpu-memory-utilization 0.40 \
    --max-model-len 4096 --enforce-eager --disable-log-requests \
    > "$LOG.serve_base" 2>&1 &

wait_url http://localhost:8103/v1/models || { echo "v6 failed"; exit 1; }
wait_url http://localhost:8104/v1/models || { echo "base failed"; exit 1; }
echo "[$(date)] both selectors ready" | tee -a "$LOG"

for CONF in "1.0 1.0 0.3" "1.0 0.5 0.3" "0.5 1.0 0.3" "1.0 1.0 0.0" "2.0 1.0 0.3" "1.0 2.0 0.3"; do
    read W1 W2 WM <<<"$CONF"
    TAG=$(echo "${W1}_${W2}_${WM}" | tr '.' '_')
    echo "[$(date)] Eval w_v6=$W1 w_base=$W2 w_maj=$WM" | tee -a "$LOG"
    $PY scripts/compute_bestofn_two_selector.py \
        "$DEV_FILE" "two_${TAG}" \
        --host_v6 http://localhost:8103 --host_base http://localhost:8104 \
        --threads 12 \
        --w_v6 $W1 --w_base $W2 --w_maj $WM --w_pok 0.5 --w_fix 0.3 \
        --out_dir eval_results 2>&1 | tee -a "$LOG.eval_${TAG}"
done

echo "[$(date)] DONE" | tee -a "$LOG"
kill_vllm